josephmayo/Mellum2-12B-A2.5B-Thinking-Abliterated
Abliterated vs JetBrains/Mellum2-12B-A2.5B-Thinking. Looks like abliteration (refusal removal): only the output projections (down_proj, embed_tokens, o_proj) changed, in 28 of 28 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 1/1, Norms 0/113, MLP 915/2773, Attention 28/112, lm_head 0/1